This is part 7, the final part, of Finance AI Field Notes. Part 6: The lonely AI architect.
Analysts will tell you the finance-AI stack is up for grabs. Job postings tell you it's already converging. Companies write job ads about the tools they've actually chosen — and in July 2026 I read about 105 live postings for people building AI inside finance functions. Extracted and tallied, the named technologies sketch a surprisingly consistent architecture.
Here's the stack, layer by layer, in the market's own words.
The model layer: named, and increasingly it's Claude
Two years ago JDs said "experience with LLMs." Now they name vendors. Kraken specifies "Claude and Anthropic APIs." Cengage lists the Anthropic API alongside OpenAI's. Wasabi wants an accountant "connecting Claude to live data." Artera goes a step further and names a product: "design and implement agentic AI solutions (e.g., Claude Code) to automate invoicing, reconciliations, and revenue recognition." And Anthropic itself is hiring a Head of Accounting AI to transform its own accounting function on Claude — the model vendor eating its own cooking.
The interesting shift isn't brand preference; it's that model choice has become an explicit architectural decision finance teams write down, the way they used to write "SAP experience required."
The integration layer: MCP appears in finance job ads now
This one would have been unthinkable in a finance JD eighteen months ago. Model Context Protocol — the standard for connecting AI models to tools and data — shows up verbatim. Kraken: "MCP or similar agent coordination layers." Wasabi: "integrating AI tools with our systems (such as NetSuite) using MCP servers." Metropolitan Commercial Bank wants "agentic workflow patterns such as MCP tool integration."
When a protocol name migrates from engineering blogs into accounting job descriptions, it's becoming plumbing. The market is converging on models talk to finance systems through standardized tool interfaces, not bespoke point integrations per use case.
The orchestration layer: n8n is the people's choice; LangGraph belongs to the banks
For the workflow engine that strings agents together, mid-market postings keep naming one tool: n8n. Kraken ("n8n or equivalent workflow engines"), Rohlik (NetSuite + Rossum + n8n), Pigment ("deploy workflows via n8n or APIs"), Pennylane (Dust, n8n). LangChain/LangGraph clusters at banks and engineering-led shops. The heavier iPaaS names (Workato, Mulesoft, Boomi) appear where an integration team already existed — Scale AI's posting is the cleanest example.
Meanwhile the RPA generation reads like the legacy line item: UiPath, Automation Anywhere and Power Automate appear mostly in the same breath as the systems being modernized. The center of gravity has moved from screen-scraping robots to API-and-LLM workflows.
One enterprise exception worth naming: at Microsoft-committed corporates, the answer is the Microsoft stack, full stop. CHANEL's Global Finance Lab builds on "Copilot for Microsoft 365, Copilot Studio, Power Automate, Power Apps." If the company runs on E5 licensing, the agent tooling follows the license.
The anchor system: NetSuite, unexpectedly central
The ERP most often named in these postings isn't SAP or Oracle Fusion — it's NetSuite: Faire, Pigment, Amplitude, Hatch, Wasabi, Cengage, Rohlik, and Kraken's finance stack all anchor on it. That makes sense once you see who's hiring builders: 200–1,500-employee companies, big enough to drown in transactions, small enough to move without a steering committee. NetSuite is where that segment lives. Around it orbit the same satellites — BlackLine, Kyriba, Coupa, Rossum, Stripe — which is exactly why the integration layer above matters so much.
The governance layer: SOX vocabulary has entered the toolchain
The most telling pattern isn't a product name at all. It's control language appearing as a technical requirement: "SOX-aligned workflows, approvals, and evidence capture" (Aon), "risk tier before work begins… audit logging confirmed" (Kraken), "validation, retries, guardrails" (Scale AI), "audit-defensible and explainable" (Aon again). Roughly half the corporate postings specify governance capabilities the way they specify programming languages. Whatever wins the finance-AI stack war, it will be something that treats an audit trail as a first-class output, not an export feature.
The stack, assembled
Put the layers together and the de facto architecture in these postings looks like this: a frontier model (increasingly Claude) → talking through standardized tool interfaces (MCP) → orchestrated by a workflow engine (n8n mid-market, LangGraph at banks, Copilot Studio in Microsoft shops) → anchored on the ERP (disproportionately NetSuite) → wrapped in SOX-grade controls that are still, today, mostly hand-built.
That last clause is the gap. Every other layer has a converging standard. The governance layer — the approval lanes, evidence capture, traceability that half these JDs demand — is the one part everyone is still building themselves, one lonely architect at a time. In any stack, the layer without a standard is where the next standard gets built.
That's the space I watch. And, in full disclosure, the space I build in.
This closes the Finance AI Field Notes series, grounded throughout in a July 2026 scan of ~105 live postings plus verified market research (Bain, L.E.K., Gartner, DSAG). I'm Andrew Rudchuk, founder of Artifi — we make the governance layer of this stack so finance teams don't hand-build it. Start from part 1, or if you're one of the companies whose job ad I quoted: I'd genuinely love to compare notes.